Lightweight checks that keep the public release runnable. They need no GPU, no
model weights, and no dataset downloads; the heavier ones skip automatically
when an optional dependency (torch, cv2, ffmpeg) is missing.
| test | what it guards |
|---|---|
test_configs.py |
every configs/*.yaml loads and resolves its ${GAE_DATA_ROOT} targets, and both a codec (gae_*) and a flow (flow_*) config exist |
test_imports.py |
the gae public facade and core research modules import, and GAE.sample keeps its public signature (needs torch; skipped otherwise) |
test_pipeline_smoke.py |
scripts/eval/smoke_test.py runs the stubbed end-to-end pipeline with no downloads (needs torch; skipped otherwise) |
test_prepare_data.py |
scripts/data/prepare_data.py emits the packed video.mp4 + meta.json schema the loaders expect, with integer-aware frame ordering (needs ffmpeg + cv2; skipped otherwise) |
pip install pytest omegaconf numpy opencv-python-headless # add torch for the import/smoke tests
pytest -q tests/.github/workflows/ci.yml runs these on a CPU-only, torch-less runner: the
config and data-prep tests execute there, while the import and smoke tests skip
cleanly (they run wherever torch is installed, e.g. a local dev checkout).